Provenance: the click that replaces the hunt
Click the total, and the invoice opens with a box drawn around the line it was read from. That is the feature.
Verification is the hidden cost
Document automation is usually sold on extraction speed. The extraction was never the slow part. The slow part is convincing yourself the extraction is right.
Without a link back to the source, that means opening the PDF, finding the right area of the right page, and reading it yourself. Per field. Per document. The tool saved you typing and handed you a search problem in exchange.
That trade is why so many teams try document AI, measure the time saved, and find it smaller than the demo implied.
What linking back actually requires
It is not enough to store which document a value came from. Every extracted field has to carry the region of the page it was read from, and that region has to survive all the way to the review surface.
So the region is data, not decoration — it is part of the extracted field, the same way the value and the score are. A pipeline that drops it somewhere between OCR and the queue cannot show you anything later.
This is the part most tools skip, and it is skipped for an understandable reason: it constrains the whole architecture rather than being a feature you add at the end.
It matters most when something is wrong
When a total looks off, provenance changes what you do next. You are not debugging the model or filing a support ticket about extraction quality. You click the field and look at the highlighted region.
Two outcomes, both fast: the document really does say that, and your expectation was wrong — or it does not, and you have found the miss and know exactly what to correct.
An unexplainable output leaves you nothing to do but distrust the whole run. A checkable one localises the problem to a field.
It outlives the review
Provenance is not only a review-time convenience. Months later, when a partner or an auditor asks where a figure came from, the answer is a click rather than an excavation.
Nothing in the books should be more than one step from the page it was read off. That is a property of the records, not a feature of the queue.